CN114565816A - Multi-modal medical image fusion method based on global information fusion - Google Patents
Multi-modal medical image fusion method based on global information fusion Download PDFInfo
- Publication number
- CN114565816A CN114565816A CN202210202366.1A CN202210202366A CN114565816A CN 114565816 A CN114565816 A CN 114565816A CN 202210202366 A CN202210202366 A CN 202210202366A CN 114565816 A CN114565816 A CN 114565816A
- Authority
- CN
- China
- Prior art keywords
- fusion
- module
- convolution
- nth
- network
- Prior art date
- Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
- Granted
Links
- 230000004927 fusion Effects 0.000 title claims abstract description 170
- 238000007500 overflow downdraw method Methods 0.000 title claims abstract description 19
- 238000006243 chemical reaction Methods 0.000 claims abstract description 4
- 238000012549 training Methods 0.000 claims abstract description 4
- 238000010586 diagram Methods 0.000 claims description 87
- 239000013598 vector Substances 0.000 claims description 26
- 230000007246 mechanism Effects 0.000 claims description 24
- 238000005070 sampling Methods 0.000 claims description 17
- 238000000034 method Methods 0.000 claims description 16
- 238000012545 processing Methods 0.000 claims description 14
- 238000010606 normalization Methods 0.000 claims description 7
- 238000011176 pooling Methods 0.000 claims description 7
- 230000003993 interaction Effects 0.000 claims description 5
- 230000008569 process Effects 0.000 claims description 4
- 238000013507 mapping Methods 0.000 claims description 3
- 239000000126 substance Substances 0.000 claims description 3
- 101100134058 Caenorhabditis elegans nth-1 gene Proteins 0.000 claims description 2
- 230000002452 interceptive effect Effects 0.000 abstract description 6
- 230000000694 effects Effects 0.000 abstract description 5
- 238000007781 pre-processing Methods 0.000 abstract description 2
- 238000010008 shearing Methods 0.000 abstract description 2
- 230000002194 synthesizing effect Effects 0.000 abstract description 2
- 230000006870 function Effects 0.000 description 12
- 238000002603 single-photon emission computed tomography Methods 0.000 description 7
- 238000013135 deep learning Methods 0.000 description 6
- 238000002600 positron emission tomography Methods 0.000 description 6
- 238000002595 magnetic resonance imaging Methods 0.000 description 5
- 238000002591 computed tomography Methods 0.000 description 4
- 210000000056 organ Anatomy 0.000 description 4
- 210000001519 tissue Anatomy 0.000 description 4
- 238000013459 approach Methods 0.000 description 3
- 238000002474 experimental method Methods 0.000 description 3
- 238000011160 research Methods 0.000 description 3
- 230000003044 adaptive effect Effects 0.000 description 2
- 238000007796 conventional method Methods 0.000 description 2
- 238000013527 convolutional neural network Methods 0.000 description 2
- 238000000354 decomposition reaction Methods 0.000 description 2
- 230000007547 defect Effects 0.000 description 2
- 238000003745 diagnosis Methods 0.000 description 2
- 238000003384 imaging method Methods 0.000 description 2
- 239000011159 matrix material Substances 0.000 description 2
- 230000002503 metabolic effect Effects 0.000 description 2
- 238000002156 mixing Methods 0.000 description 2
- 238000012360 testing method Methods 0.000 description 2
- 230000009466 transformation Effects 0.000 description 2
- NAWXUBYGYWOOIX-SFHVURJKSA-N (2s)-2-[[4-[2-(2,4-diaminoquinazolin-6-yl)ethyl]benzoyl]amino]-4-methylidenepentanedioic acid Chemical compound C1=CC2=NC(N)=NC(N)=C2C=C1CCC1=CC=C(C(=O)N[C@@H](CC(=C)C(O)=O)C(O)=O)C=C1 NAWXUBYGYWOOIX-SFHVURJKSA-N 0.000 description 1
- 206010028980 Neoplasm Diseases 0.000 description 1
- 210000003484 anatomy Anatomy 0.000 description 1
- 230000009286 beneficial effect Effects 0.000 description 1
- 230000017531 blood circulation Effects 0.000 description 1
- 210000000988 bone and bone Anatomy 0.000 description 1
- 230000008859 change Effects 0.000 description 1
- 230000000052 comparative effect Effects 0.000 description 1
- 230000000295 complement effect Effects 0.000 description 1
- 238000013136 deep learning model Methods 0.000 description 1
- JXSJBGJIGXNWCI-UHFFFAOYSA-N diethyl 2-[(dimethoxyphosphorothioyl)thio]succinate Chemical compound CCOC(=O)CC(SP(=S)(OC)OC)C(=O)OCC JXSJBGJIGXNWCI-UHFFFAOYSA-N 0.000 description 1
- 201000010099 disease Diseases 0.000 description 1
- 208000037265 diseases, disorders, signs and symptoms Diseases 0.000 description 1
- 238000005516 engineering process Methods 0.000 description 1
- 238000011156 evaluation Methods 0.000 description 1
- 238000000605 extraction Methods 0.000 description 1
- 239000007943 implant Substances 0.000 description 1
- 230000006872 improvement Effects 0.000 description 1
- 238000012880 independent component analysis Methods 0.000 description 1
- 238000002075 inversion recovery Methods 0.000 description 1
- 238000011835 investigation Methods 0.000 description 1
- 239000007788 liquid Substances 0.000 description 1
- 238000005259 measurement Methods 0.000 description 1
- 230000004060 metabolic process Effects 0.000 description 1
- 230000008447 perception Effects 0.000 description 1
- 239000011541 reaction mixture Substances 0.000 description 1
- 210000004872 soft tissue Anatomy 0.000 description 1
Images
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F18/00—Pattern recognition
- G06F18/20—Analysing
- G06F18/25—Fusion techniques
- G06F18/253—Fusion techniques of extracted features
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F18/00—Pattern recognition
- G06F18/20—Analysing
- G06F18/21—Design or setup of recognition systems or techniques; Extraction of features in feature space; Blind source separation
- G06F18/214—Generating training patterns; Bootstrap methods, e.g. bagging or boosting
- G06F18/2155—Generating training patterns; Bootstrap methods, e.g. bagging or boosting characterised by the incorporation of unlabelled data, e.g. multiple instance learning [MIL], semi-supervised techniques using expectation-maximisation [EM] or naïve labelling
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/04—Architecture, e.g. interconnection topology
- G06N3/045—Combinations of networks
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/08—Learning methods
- G06N3/088—Non-supervised learning, e.g. competitive learning
-
- Y—GENERAL TAGGING OF NEW TECHNOLOGICAL DEVELOPMENTS; GENERAL TAGGING OF CROSS-SECTIONAL TECHNOLOGIES SPANNING OVER SEVERAL SECTIONS OF THE IPC; TECHNICAL SUBJECTS COVERED BY FORMER USPC CROSS-REFERENCE ART COLLECTIONS [XRACs] AND DIGESTS
- Y02—TECHNOLOGIES OR APPLICATIONS FOR MITIGATION OR ADAPTATION AGAINST CLIMATE CHANGE
- Y02D—CLIMATE CHANGE MITIGATION TECHNOLOGIES IN INFORMATION AND COMMUNICATION TECHNOLOGIES [ICT], I.E. INFORMATION AND COMMUNICATION TECHNOLOGIES AIMING AT THE REDUCTION OF THEIR OWN ENERGY USE
- Y02D10/00—Energy efficient computing, e.g. low power processors, power management or thermal management
Landscapes
- Engineering & Computer Science (AREA)
- Theoretical Computer Science (AREA)
- Physics & Mathematics (AREA)
- Data Mining & Analysis (AREA)
- Life Sciences & Earth Sciences (AREA)
- Artificial Intelligence (AREA)
- General Physics & Mathematics (AREA)
- General Engineering & Computer Science (AREA)
- Evolutionary Computation (AREA)
- Computer Vision & Pattern Recognition (AREA)
- Computational Linguistics (AREA)
- Bioinformatics & Computational Biology (AREA)
- Bioinformatics & Cheminformatics (AREA)
- Health & Medical Sciences (AREA)
- Biomedical Technology (AREA)
- Biophysics (AREA)
- Evolutionary Biology (AREA)
- General Health & Medical Sciences (AREA)
- Molecular Biology (AREA)
- Computing Systems (AREA)
- Mathematical Physics (AREA)
- Software Systems (AREA)
- Image Analysis (AREA)
- Image Processing (AREA)
Abstract
The invention discloses a multi-modal medical image fusion method based on global information fusion, which comprises the following steps: 1. carrying out color space conversion and image shearing preprocessing on the original medical images of a plurality of modalities; 2. establishing a modal branch network which interacts through a fusion module in a plurality of scales, and establishing a fusion module consisting of a Transformer to combine multi-modal characteristic information; 3. establishing a reconstruction module, and synthesizing a fusion image from multi-scale multi-modal characteristics; 4. training and evaluating the model on a public data set; 4. and realizing a medical image fusion task by using the trained model. The invention can fully fuse the multi-modal semantic information through the Transformer fusion module and the interactive modal branch network, realizes the fine-grained fusion effect, well retains the structure and texture information of the original image, and improves the mosaic phenomenon caused by the low-resolution medical image.
Description
Technical Field
The invention relates to the technical field of image fusion, in particular to a medical image fusion technology based on deep learning.
Background
Medical images can assist doctors in better understanding human structures and tissues, and are widely used in clinical applications such as disease diagnosis, treatment planning, surgical guidance, and the like. Due to the difference of imaging mechanisms, the attention degree of medical images of different modalities to human organs and tissues is different. The medical image of single modality often can not provide comprehensive and sufficient information, and doctors often need to observe a plurality of images simultaneously to accurately judge the state of an illness, which inevitably brings certain difficulty to diagnosis. Due to the limitations of single modality medical images, multi-modality medical image fusion is a field in which research is very necessary. The multi-modal medical image fusion is to synthesize an image by synthesizing important information of medical images of different modalities in the same scene.
Generally, medical images can be divided into anatomical images and functional images. The anatomical image has high spatial resolution, can clearly image the anatomical structure of organs, but cannot display the function change of human metabolism, such as Computed Tomography (CT) and Magnetic Resonance imaging (MR); functional images, on the contrary, are very functional and metabolic displays, but have low resolution and do not accurately describe anatomical details of organs, such as Positron Emission Tomography (PET) and Single-Photon Emission Computed Tomography (Single-Photon Emission Computed Tomography). Even if CT and MR are both anatomical images and PET and SPECT are both functional images, the information they are interested in is different. CT mainly reflects the position information of human bones and implants, while MR is mainly used to provide clear detailed information to soft tissues and other parts. MR also contains multiple modalities, focusing on sub-regions of different properties, and the commonly used modalities are T1 weighted (denoted as T1), contrast enhanced T1 weighted (denoted as T1c), T2 weighted (denoted as T2), and liquid decay inversion recovery pulse (denoted as FLAIR). PET primarily reflects tumor function and metabolic information, whereas SPECT primarily provides organ and tissue blood flow information.
Most multi-modality medical image fusion methods can be summarized in three processes: and (5) extracting, fusing and reconstructing the characteristics. In order to realize medical image fusion, various algorithms have been proposed by scholars at home and abroad for over three decades, and generally speaking, the methods can be divided into two major categories: a conventional fusion method and a fusion method based on deep learning.
In the traditional medical image fusion framework, researchers propose a plurality of decomposition modes or transformation modes to extract the characteristics of a source image, then select a certain fusion strategy to fuse the characteristics, and finally, perform inverse transformation on the fused characteristics to reconstruct a fused image. Conventional methods can be divided into four categories according to the way of feature extraction: (1) sparse representation-based methods; (2) methods based on multi-scale decomposition, such as pyramids and wavelets; (3) subspace-based methods, such as independent component analysis; (4) a method based on salient features. The traditional medical image fusion method has achieved good fusion effect, but has some defects, and further improvement of fusion performance is limited. First, the fusion performance of the conventional approach relies heavily on artificially defined features, which may limit the generalization of the approach to other fusion tasks. Second, different features may require different fusion strategies to function. Third, for the fusion method based on sparse representation, the dictionary learning is relatively time-consuming, so that it takes much time to synthesize a fusion image.
In recent years, a method based on deep learning becomes a new research hotspot in the field of image fusion, a deep learning model represented by a Convolutional Neural Network (CNN) and a Generative adaptive Network (generic adaptive Network) is successfully applied to image fusion problems of multi-focus, infrared, visible light and the like, characteristics and fusion strategies do not need to be defined artificially, and the method has the advantages over the traditional fusion method. However, because a reference image of a fusion result cannot be constructed for supervised learning and the complex diversity of human structures and tissues causes that the imaging characteristics of each modality are not easy to quantitatively describe and other factors, the current medical image fusion method based on deep learning has relatively few researches and is still in a starting stage.
Investigations have found that existing medical image fusion methods typically use either a human-defined fusion strategy or a convolution-based network to fuse multimodal image features. However, these fusion strategies do not efficiently extract global semantic information of multimodal images. In addition, the current medical image fusion method based on deep learning has the problems of insufficient and imprecise utilization of multi-modal image information. Most approaches use multimodal images in a simplistic way, the most common way being to stack the original different modality images (or the respective extracted underlying features) in channel dimensions and then input them directly into the network model for fusion.
Disclosure of Invention
The invention provides a medical image fusion method based on global information fusion to overcome the defects of the prior art, so that global information with multi-modal characteristics can be combined through a self-attention mechanism, and information of different modes can be utilized to the maximum extent through an interactive mode branch network, and a high-quality medical image fusion effect is realized.
In order to achieve the purpose, the invention adopts the following technical scheme:
the invention relates to a multi-modal medical image fusion method based on global information fusion, which is characterized by comprising the following steps:
step one, obtaining original medical images of M different modes and carrying out conversion of YCbCr color space to obtain Y channel images { I) of all modes1,...,Im...,IM}; wherein, ImA Y-channel image representing the mth modality, M ∈ {1, 2.., M }; y-channel image for all modalities { I1,...,Im...,IMCutting the image to obtain the image block set S of all the modes1,...,Sm...,SMIn which S ismA set of image blocks representing an mth modality;
step two, constructing a fusion network Transfusion, comprising the following steps: m modal branch networks, N fusion modules and a reconstruction module; and sets the image blocks of all the modalities { S1,...,Sm...,SMInputting into the fusion network TransFusion:
step 2.1, constructing the M modal branch networks and the N fusion modules:
step 2.1.1, M modal branch networks are constructed:
the mth modal branch network in the M modal branch networks is composed of N convolution modules, and the N convolution modules are respectively marked as ConvBlockm1,...,ConvBlockmn,...,ConvBlockmNWherein, ConvBlockmnAn nth convolution module representing an mth modal branching network, N ∈ {1, 2.., N };
when n is 1, the nth convolution module convBlock of the mth modal branching networkmnFrom XmnTwo-dimensional convolution layers;
when N is 2, 3.. times.n, the nth convolution module ConvBlock of the mth modal branching networkmnFrom a maximum pooling layer and XmnTwo-dimensional convolution layers;
the convolution kernel size of the x two-dimensional convolution layer of the n convolution module of the m modal branching network is ksmnx×ksmnxThe number of convolution kernels is knmnx,x∈{1,2,...,Xmn};
Step 2.1.2, constructing N fusion modules:
any nth fusion module of the N fusion modules is a Transformer network and consists of L self-attention mechanism modules; the first of the L self-attention mechanism modules comprises: 1 multi-head attention layer, 2 layers of normalization and 1 full-connection layer;
step 2.2, gathering image blocks of all modes { S1,...,Sm...,SMInputting into M modal branch networks, and performing information fusion through N fusion modules:
when n is 1, the set S of image blocks of the mth modalitymAn nth convolution module ConvBlock input to the mth modal branching networkmnX of (2)mnOutputting characteristic diagram after two-dimensional convolution layerWherein Hn、Wn、DnRespectively representing the output characteristic diagram of the mth modal branch network in the nth convolution moduleHigh, wide, number of channels; thereby obtaining the output characteristic diagram of the M modal branch networks in the nth convolution module
Output characteristic diagram of the M modal branch networks in the nth convolution moduleAfter the n-th fusion module processes, outputting a characteristic diagramWherein the content of the first and second substances,friend the mth characteristic diagram output by the nth fusion module;
the mth feature map output by the nth fusion moduleFeature map output by the nth convolution module of the mth modal branching networkAdding to obtain the characteristic diagram of the mth modal branch network after the nth convolution module interactionThereby obtaining the characteristic diagram of the M modal branch networks after the n convolution module interaction
When N is 2,3, the characteristic diagram of the mth modal branch network after the interaction of the nth-1 convolution modulesInputting the data into the maximum pooling layer of the nth convolution module of the mth modal branch network for down-sampling to obtain a feature map of the mth modal branch network after down-sampling in the nth convolution moduleThe feature map after down samplingInputting the data into the 1 st two-dimensional convolution layer of the nth convolution module of the mth modal branch network, and sequentially passing through XmAfter the two-dimensional convolution layer is processed, a characteristic diagram is outputThereby obtaining a characteristic diagram output by the nth convolution module of the M modal branch networksThe characteristic diagram is combinedThrough the n fusionAfter the module is processed, the characteristic diagram is outputThe mth feature map output by the nth fusion moduleCharacteristic diagram output by nth convolution module of mth modal branch networkAdding to obtain an added characteristic diagramThereby obtaining M added feature mapsFurther obtaining a characteristic diagram output by the Nth fusion module
Step 2.3, the reconstruction module is composed of an N-level convolution network; and outputting the feature maps of the N fusion modulesInputting the image data into the reconstruction module to obtain a preliminary fusion image F':
step 2.3.1, all characteristic graphs output by the n fusion moduleAdding to obtain a fused feature map phin(ii) a Thereby obtaining a fusion characteristic diagram { phi ] of N fusion modules1,...,Φn,...,ΦN};
Step 2.3.2, a reconstruction module formed by N-level convolution networks is constructed, and the fusion characteristic diagram phi of the N-th fusion module is usednThe nth stage convolution network input to the reconstruction module:
when n is 1The nth stage convolutional network comprises: b isnConvolution module RConvBlockn1,...,RConvBlocknb,...,
When N is 2, 3.. times.n, the nth stage convolutional network includes: b isnConvolution module RConvBlockn1,...,RConvBlocknb,...,And Bn+1 upsampling layer UpSamplen1,...,UpSamplenb,...,The b convolution module RConvBlock of the nth stage convolution networknbConsists of Y two-dimensional convolution layers, B is in the form of {1,2n};
When n is 1 and b is 1, fusing the feature map phi of the n-th fusion modulenA b-th convolution module RConvBlock input to the n-th stage convolution networknbAnd outputting a characteristic diagram phi Rnb;
When N is 2,3,.., N and b is 1, the fusion feature map Φ of the N-th fusion module is comparednUpSample of the b-th upsampling layer input to the n-th convolutional networknbAfter up-sampling processing is carried out, an up-sampled characteristic diagram phi U is outputnb(ii) a Thereby obtaining the characteristic diagram { phi U after the up-sampling of the convolution networks from the 2 nd level to the N-1 th level2b,...,ΦUnb,...,ΦUNb};
When N is 2,3, thenThen, the fusion characteristic diagram phi of the n-th fusion module is combinednOutput characteristic diagram { phi R of first b-1 convolution modules of nth-level convolution networkn1,...,ΦRn(b-1)B characteristic graphs after up-sampling of the first b of the (n + 1) th level convolution network (phi U)(n+1)1,...,ΦU(n+1)bSplicing to obtain a spliced characteristic diagram; inputting the spliced feature map to the nth stageThe b-th convolution module RConvBlock of the convolution networknbAnd outputting the output characteristic diagram phi R of the b convolution module of the nth-level convolution networknb(ii) a Thereby obtaining the B-th of the 1 st convolution network1Output characteristic diagram of convolution module
The B-th of the 1 st level convolution network1Output characteristic diagram of convolution moduleAfter processing of a convolution layer, obtaining a primary fusion image F';
step three, constructing a loss function and training a network to obtain an optimal fusion model:
step 3.1, respectively calculating image block sets { S of all modes1,...,Sm...,SMEntropy of each image block set in the image block set is obtained, and a corresponding entropy value { e }is obtained1,...,em...,eMIn which emEntropy values of a set of image blocks representing an mth modality;
step 3.2, for the entropy value { e1,...,em...,eMRespectively carrying out normalization processing to obtain image block sets (S) of all modes1,...,Sm...,SMWeight of [ omega ]1,...,ωm,...,ωMIn which ω ismWeights representing a set of image blocks of the mth modality;
step 3.3, constructing a total Loss function Loss by using the formula (1):
in the formula (1), Lssim(SmF') represents the set S of image blocks of the m-th modalitymA structural similarity loss function with the preliminary fused image F';
step 3.4, an optimizer is used for carrying out minimum solving on the total Loss function Loss, so that all parameters in the fusion network Transfusion are optimized, and an optimal fusion model is obtained;
step four, utilizing the optimal fusion model to perform Y-channel image { I) of all modes1,I2,...,IMProcessing and outputting a primary fusion image F'; the preliminary fused image F' is converted into an RGB color space, thereby obtaining a final fused image F.
The multi-modal medical image fusion method based on the global information fusion is also characterized in that the nth fusion module in the step 2.2 is processed according to the following processes:
step 2.2.1, the characteristic diagram of the N-th convolution module output by the n-th fusion module to the M modal branch networksSplicing and leveling to obtain the size of (M H)n*Wn)×DnThe flattened feature vectors of (a); adding the flattened feature vector and a trainable vector with the same size to obtain the feature vector containing the position information of the nth fusion module
Step 2.2.2, when l is 1, the 1 st self-attention mechanism module of the nth fusion module uses the feature vectorAfter linear mapping, three matrixes Q are obtainednl,Knl,Vnl(ii) a Recalculate Qnl,Knl,VnlMultiple head attention results Z in betweennl(ii) a The multiple head attention result ZnlInputting the data into a full connection layer of the ith self-attention mechanism module of the nth fusion module, and obtaining an output sequence vector of the ith self-attention mechanism module of the nth fusion module
When l is 2,3,.. times.N, the output sequence vector of the l-1 self-attention mechanism module of the N-th fusion module is usedInputting the input into the first self-attention mechanism module of the nth fusion module, and obtaining the output sequence vector of the first self-attention mechanism module of the nth fusion moduleThereby obtaining the output sequence vector of the L-th self-attention mechanism module of the n-th fusion module
Step 2.2.3, outputting the output sequence vectorDividing into M modes, and deforming the size of each mode into Hn×Wn×DnTo obtain an output characteristic diagram
Compared with the prior art, the invention has the beneficial effects that:
1. the invention provides an unsupervised anatomical and functional image fusion method. The method introduces a Transformer structure as a fusion strategy, and the Transformer combines the global information of the multi-modal medical image and sufficiently fuses the semantic information of the multi-modal image by utilizing a self-attention mechanism contained in the Transformer, so that a fine-grained fusion effect is realized. The invention not only well retains the structure and texture information in the anatomical image, but also improves the mosaic phenomenon caused by low resolution in the functional image.
2. The present invention proposes modal branching networks that interact on multiple scales. The network can extract multi-scale complementary features of each modal image, and fully utilizes multi-modal image information. The interactive branched network enhances the anatomical and functional image fusion effect.
Drawings
FIG. 1 is a schematic flow chart of the method of the present invention, wherein "ks × ks, kn" represents a convolution layer with a kernel size of ks × ks and a number of kernels of kn;
fig. 2 is a structural diagram of a modal branching network and a convergence module provided in an embodiment of the present invention;
fig. 3 is a structural diagram of a reconstruction module according to an embodiment of the present invention.
Detailed Description
In this embodiment, a multi-modal image fusion method based on global information fusion, as shown in fig. 1, includes the following steps:
step one, obtaining M original medical images of different modalities and carrying out preprocessing of color space conversion and image shearing to obtain a preprocessed image block set { S) of all modalities1,S2,...,SMIn which S ismA set of image blocks representing the mth modality, M ∈ {1, 2.., M }:
step 1.1, acquiring original medical images of a plurality of modalities required by an experiment from a Harvard medical image data set website (http:// www.med.harvard.edu/AANLIB/home. html); the present embodiment collects medical images of M-2 modalities from this public dataset, including 279 pairs of MR-T1 and PET images and 318 pairs of MR-T2 and SPECT images, where MR-T1 and MR-T2 are anatomical images in grayscale, the number of channels is 1, PET and SPECT are functional images of RGB color space, and the number of channels is 3;
step 1.2, converting the image of the RGB color space into the YCbCr space according to the formula (1):
in formula (1), R, G, B are three channels of RGB color space, Y is a luminance channel, and Cb and Cr are two color channels, respectively;
step (ii) of1.3, in order to expand the number of samples, cutting the gray level image and the image of the Y channel into image blocks to obtain an image block set { S) of all modes1,S2,...,SM}; in this embodiment, the size of the clipped image block is 64 × 64;
step two, constructing a fusion network Transfusion, comprising the following steps: m modal branch networks, N fusion modules and a reconstruction module; and sets the image blocks of all the modalities { S1,...,Sm...,SMInputting into a fusion network TransFusion:
step 2.1, constructing M modal branch networks and N fusion modules:
step 2.1.1, constructing M modal branch networks:
the mth modal branch network in the M modal branch networks is composed of N convolution modules, and the N convolution modules are respectively marked as ConvBlockm1,...,ConvBlockmn,...,ConvBlockmNWherein, ConvBlockmnAn nth convolution module representing an mth modal branching network, N ∈ {1, 2.., N };
when n is equal to 1, the nth convolution module convBlock of the mth modal branching networkmnFrom XmnTwo-dimensional convolution layers;
when N is 2, 3.. times.n, the nth convolution module ConvBlock of the mth modal branching networkmnFrom a maximum pooling layer and XmnTwo-dimensional convolution layers;
the convolution kernel size of the x two-dimensional convolution layer of the n convolution module of the m modal branching network is ksmnx×ksmnxThe number of convolution kernels is knmnx,x∈{1,2,...,Xmn};
In this embodiment, N is 4, the kernel size of all the largest pooling layers is 2 × 2, the step size is 2, and X ismn、ksmnx、knmnxAs shown in fig. 2;
step 2.1.2, constructing N fusion modules:
any nth fusion module of the N fusion modules is a Transformer network and consists of L self-attention mechanism modules; the first of the L self-attention mechanism modules comprises: 1 multi-head attention layer, 2 layers of normalization and 1 full-connection layer; in the present embodiment, L ═ 1;
step 2.2, gathering image blocks of all modes { S1,...,Sm...,SMInputting into M modal branch networks, and performing information fusion through N fusion modules:
when n is 1, the set S of image blocks of the mth modalitymThe nth convolution module ConvBlock input to the mth modal branching networkmnX of (2)mnOutputting characteristic diagram after two-dimensional convolution layerWherein Hn、Wn、DnRespectively representing the output characteristic diagram of the mth modal branch network at the nth convolution moduleHigh, nest, channel count; thereby obtaining the output characteristic diagram of the M modal branch networks in the nth convolution moduleIn this example, (H)1,W1,D1)=(64,64,64),(H2,W2,D2)=(32,32,128),(H3,W3,D3)=(16,16,256),(H4,W4,D4)=(8,8,512);
Output characteristic diagram of M modal branch networks in nth convolution moduleAfter the n-th fusion module is processed according to the formula (2), a characteristic diagram is outputWherein, the first and the second end of the pipe are connected with each other,an mth feature map representing an output of the nth fusion module:
in the formula (2), a TransformernRepresenting the nth fusion module and realizing the following steps:
step 2.2.1, the characteristic diagram of the n-th convolution module output by the n-th fusion module to the M modal branch networksSplicing and leveling to obtain the size of (M H)n*Wn)×DnThe flattened feature vectors of (a); adding the flattened feature vector with a trainable vector with the same size to obtain the feature vector containing the position information of the nth fusion module
Step 2.2.2, the feature vector is processed by the self-attention mechanism module of the n fusion moduleLinear mapping to three matrices, Qn,Kn,Vn:
Step 2.2.3, mixing Qn,Kn,VnAre respectively divided into h heads to obtaini ∈ {1, 2., h }, and then the multi-head attention is calculated according to equation (4) -equation (6), resulting in a result Z:
in the formula (5), Concat represents a splicing operation,is a trainable matrix; in formula (6), LayerNorm indicates layer normalization;
step 2.2.4, result Z of multi-head attentionnInputting the data into a full connection layer to obtain an output sequence vector of the nth fusion module
In formula (7), MLP represents a fully connected layer;
step 2.2.5, mixingDividing into M modes, and deforming the size of each mode into Hn×Wn×DnObtaining an output characteristic diagramWherein the content of the first and second substances,an mth feature map representing an output of the nth fusion module;
the mth feature map output by the nth fusion moduleCharacteristic diagram output by nth convolution module of mth modal branch networkAdding to obtain a characteristic diagram of the mth modal branch network after the nth convolution module is interactedThereby obtaining the characteristic diagram of the M modal branch networks after the n convolution module interaction
When N is 2,3,.. times, N, the characteristic diagram of the mth modal branch network after the interaction of the (N-1) th convolution module is obtainedInputting the data into the maximum pooling layer of the nth convolution module of the mth modal branch network for down-sampling to obtain a feature map of the mth modal branch network after down-sampling in the nth convolution moduleThe feature map after down samplingTransporting into the 1 st two-dimensional convolution layer of the nth convolution module of the mth modal branch network, and sequentially passing through XmnAfter the two-dimensional convolution layer is processed, a characteristic diagram is outputThereby obtaining M modal branched netsCharacteristic diagram output by nth convolution module of networkWill feature mapAfter the processing of the n-th fusion module, a characteristic diagram is outputThe mth feature map output by the nth fusion moduleCharacteristic diagram output by nth convolution module of mth modal branch networkAdding to obtain an added characteristic diagramThereby obtaining M added feature mapsFurther obtaining a characteristic diagram output by the Nth fusion module
Step 2.3, constructing a reconstruction module consisting of N-level convolution networks, and fusing output characteristic graphs of the N fusion modulesInputting the image into a reconstruction module to obtain a preliminary fusion image F':
step 2.3.1, all characteristic graphs output by the n fusion moduleAdding to obtain a fused feature map phin(ii) a Thereby obtaining N fusion modulesFused feature map of blocks Φ1,...,Φn,...,ΦN};
Step 2.3.2, a reconstruction module formed by N-level convolution networks is constructed, and the fusion characteristic diagram phi of the N-th fusion module is usednThe nth level convolution network input to the reconstruction module:
when n is 1, the nth stage convolutional network includes: b isnConvolution module RConvBlockn1,RConvBlockn2,...,
When N is 2, 3.. times.n, the nth stage convolutional network includes: b isnConvolution module RConvBlockn1,RConvBlockn2,...,And Bn+1 upsampling layer UpSamplen1,UpSamplen2,...,The b convolution module RConvBlock of the nth stage convolution networknbConsists of Y two-dimensional convolution layers, B is in the form of {1,2n};
When n is 1 and b is 1, fusing characteristic diagram phi of the n-th fusing modulenThe b convolution module RConvBlock input to the nth stage convolution networknbAnd outputting a characteristic diagram phi Rnb;
When N is 2,3, the. N, and b is 1, the fusion feature map Φ of the N-th fusion module is comparednInput to the b-th upsampling layer UpSample of the nth stage convolutional networknbAfter up-sampling processing is carried out, an up-sampled characteristic diagram phi U is outputnb(ii) a Thereby obtaining the characteristic diagram { phi U after the up-sampling of the convolution networks from the 2 nd level to the N-1 th level2b,...,ΦUnb,...,ΦUNb};
When N is 2,3, thenThen, the fusion characteristic diagram phi of the n-th fusion module is combinednN-th order convolutionOutput characteristic diagram { phi R of first b-1 convolution modules of networkn1,...,ΦRn(b-1)B characteristic graphs after up-sampling of the first b of the (n + 1) th level convolution network (phi U)(n+1)1,...,ΦU(n+1)bSplicing to obtain a spliced characteristic diagram; inputting the spliced feature map into a b-th convolution module RConvBlock of an nth-stage convolution networknbAnd outputting the output characteristic diagram phi R of the b convolution module of the nth-level convolution networknb(ii) a Thereby obtaining the B-th of the 1 st convolution network1Output characteristic diagram of convolution module
Convolution network B of level 11Output characteristic diagram of convolution moduleAfter processing of a convolution layer, obtaining a primary fusion image F';
in this embodiment, the reconstruction module is shown in fig. 3, where Y is 2 and B is1=3,B2=2,B3=1,B4=0;
Step three, constructing a loss function and training a network to obtain an optimal fusion model:
step 3.1, calculating image block sets { S of all the modalities according to the formulas (8) to (9)1,S2,...,SMEntropy of each image block set in the image block set is obtained, and a corresponding entropy value { e }is obtained1,e2,...,eMIn which emEntropy values of a set of image blocks representing the mth modality:
em=Entropy(Sm) (8)
in the formula (9), plProbability of being the l-th gray value;
step 3.2, entropy value { e }1,e2,...,eMRespectively carrying out normalization processing to obtain image block sets (S) of all modes1,S2,...,SMWeight of { omega } ω1,ω2,...,ωMIn which ω ismWeight of set of image blocks representing mth modality:
in the formula (10), η is a temperature parameter; in the present embodiment, η ═ 1;
step 3.3, constructing a total Loss function Loss by using the formula (11):
Loss=ω1Lssim(S1,F′)+ω2Lssim(S2,F′) (11)
Lssim(Sj,F′)=1-SSIM(Sj,F′) (12)
in the formula (11), Lssim(SmF') represents the set S of image blocks of the m-th modalitymA structural similarity loss function with the preliminary fused image F'; in formula (12), SSIM is a structural similarity function;
step 3.4, adopting an AdamW optimizer to carry out minimum solution on the Loss function Loss, thereby optimizing all parameters in the fusion network Transfusion and obtaining an optimal fusion model;
step four, utilizing the optimal fusion model to perform Y-channel image or gray level image { I) of all modes1,I2,...,IMProcessing and outputting a primary fusion image F'; splicing the preliminary fusion image F 'with Cb and Cr channels and converting the preliminary fusion image F' into an RGB color space to obtain a final fusion image F;
step five, evaluating the performance of the invention:
in specific implementation, the present invention is compared with the conventional methods CSMCA and the deep learning methods DDcGAN and EMFusion. In addition, for explaining the Transformer-based fusion module and the interactive modal branching network in the inventionTwo comparative experiments were set up. The first experiment removed the Transformer fusion module and the second replaced the interactive modal branching network with a weight-shared modal branching network. Boundary-based similarity measurement Q using mutual information, mean gradientAB/FVisual perception index QCVMutual information, average gradient, Q as evaluation indexAB/FThe larger, QCVThe smaller the quality of the fused image. The average fusion quality for the 30 pairs of MR-T1 and PET test images and the 30 pairs of MR-T2 and SPECT test images is as follows:
TABLE 1. fusion Performance of different methods
The experimental result shows that the invention has mutual information, average gradient and QAB/F、QCVThe four indexes are all optimal. The Transformer fusion module of the invention helps to enhance mutual information of 5.10% -10.02%, average gradient of 2.59% -5.28%, Q of 3.04% -4.36%AB/F1.43% -12.66% of QCV(ii) a The inventive interactive modal branching network helps to enhance mutual information of 18.39% -19.91%, average gradient of 1.06% -6.69%, Q of 7.68% -11.02%AB/F27.69% -62.22% of QCV。
Claims (2)
1. A multi-modal medical image fusion method based on global information fusion is characterized by comprising the following steps:
step one, obtaining original medical images of M different modes and carrying out conversion of YCbCr color space to obtain Y channel images { I) of all modes1,…,Im…,IM}; wherein, ImA Y-channel image representing the mth modality, M ∈ {1,2, …, M }; y-channel image for all modalities { I1,…,Im…,IMCutting the image to obtain the image block set S of all the modes1,…,Sm…,SMIn which SmDiagram representing the m-th modalityA set of blocks;
step two, constructing a fusion network Transfusion, which comprises the following steps: m modal branch networks, N fusion modules and a reconstruction module; and sets the image blocks of all the modalities { S1,…,Sm…,SMInputting into the fusion network TransFusion:
step 2.1, constructing the M modal branch networks and the N fusion modules:
step 2.1.1, constructing M modal branch networks:
the mth modal branch network in the M modal branch networks is composed of N convolution modules, and the N convolution modules are respectively recorded as ConvBlockm1,…,COnvBlockmn,…,ConvBlockmNWherein, ConvBlockmnAn nth convolution module representing an mth modal branching network, N ∈ {1,2, …, N };
when n is 1, the nth convolution module convBlock of the mth modal branching networkmnFrom XmnTwo-dimensional convolution layers;
when N is 2,3, …, N, the nth convolution module ConvBlock of the mth modal branching networkmnFrom a maximum pooling layer and XmnTwo-dimensional convolution layers;
the convolution kernel size of the x two-dimensional convolution layer of the n convolution module of the m modal branching network is ksmnx×ksmnxThe number of convolution kernels is knmnx,x∈{1,2,…,Xmn};
Step 2.1.2, constructing N fusion modules:
any nth fusion module of the N fusion modules is a Transformer network and consists of L self-attention mechanism modules; the first of the L self-attention mechanism modules comprises: 1 multi-head attention layer, 2 layers of normalization and 1 full connection layer;
step 2.2, gathering image blocks of all modes { S1,…,Sm…,SMInputting into M modal branch networks, and performing information fusion through N fusion modules:
when n is 1, the map of the mth modeSet of blocks SmThe nth convolution module ConvBlock input to the mth modal branching networkmnX of (2)mnOutputting characteristic diagram after two-dimensional convolution layerWherein Hn、Wn、DnRespectively representing the output characteristic diagram of the mth modal branch network in the nth convolution moduleHigh, wide, number of channels; thereby obtaining the output characteristic diagram of the M modal branch networks in the nth convolution module
Output characteristic diagram of the M modal branch networks in the nth convolution moduleAfter the n-th fusion module processes, outputting a characteristic diagramWherein the content of the first and second substances,an mth feature map representing an output of the nth fusion module;
the mth feature map output by the nth fusion moduleFeature map output by the nth convolution module of the mth modal branching networkAdding to obtain the characteristic diagram of the mth modal branch network after the nth convolution module interactionThereby obtaining the characteristic diagram of the M modal branch networks after the n convolution module interactionWhen N is 2,3, …, N, the characteristic diagram of the mth modal branching network after the interaction of the nth-1 convolution moduleInputting the data into the maximum pooling layer of the nth convolution module of the mth modal branch network for down-sampling to obtain a feature map of the mth modal branch network after down-sampling in the nth convolution moduleThe feature map after down samplingInputting the data into the 1 st two-dimensional convolution layer of the nth convolution module of the mth modal branch network, and sequentially passing through XmnAfter the two-dimensional convolution layer is processed, a characteristic diagram is outputThereby obtaining a characteristic diagram output by the nth convolution module of the M modal branch networksThe characteristic diagram is combinedAfter the processing of the n-th fusion module, a characteristic diagram is outputThe mth feature map output by the nth fusion moduleCharacteristic diagram output by nth convolution module of mth modal branch networkAdding to obtain an added characteristic diagramThereby obtaining M added feature mapsFurther obtaining a characteristic diagram output by the Nth fusion module
Step 2.3, the reconstruction module is composed of an N-level convolution network; and outputting the feature maps of the N fusion modulesInputting the image data into the reconstruction module to obtain a preliminary fusion image F':
step 2.3.1, all characteristic graphs output by the n fusion moduleAdding to obtain a fused feature map phin(ii) a Thereby obtaining a fusion characteristic diagram { phi ] of N fusion modules1,…,Φn,…,ΦN};
Step 2.3.2, a reconstruction module formed by N-level convolution networks is constructed, and the fusion characteristic diagram phi of the N-th fusion module is usednThe nth stage convolution network input to the reconstruction module:
When N is 2,3, …, N, the nth stage convolutional network comprises: b isnA convolution module And Bn+1 upsampling layers The b convolution module RConvBlock of the nth stage convolution networknbConsists of Y two-dimensional convolution layers, B ∈ {1,2, …, Bn};
When n is 1 and b is 1, fusing the feature map phi of the n-th fusion modulenA b-th convolution module RConvBlock input to the n-th stage convolution networknbAnd outputting a characteristic diagram phi Rnb;
When N is 2,3, …, N and b is 1, fusing the feature map phi of the N-th fusing modulenUpSample of the b-th upsampling layer input to the n-th convolutional networknbAfter up-sampling processing is carried out, an up-sampled characteristic diagram phi U is outputnb(ii) a Thereby obtaining the up-sampled characteristic diagram { phi U of the convolution network from the 2 nd level to the N-1 st level2b,…,ΦUnb,…,ΦUNb};
When N is 2,3, …, N and B is 2,3, …, BnThen, the fusion characteristic diagram phi of the n-th fusion module is combinednOutput characteristic diagram { phi R of the first b-1 convolution modules of the nth-level convolution networkn1,…,ΦRn(b-1) B characteristic graphs after up-sampling of the first b of the (n + 1) th level convolution network (phi U)(n+1)1,…,ΦU(n+1)bAfter splicing, obtainingSplicing the characteristic graphs; inputting the spliced feature map into a b-th convolution module RConvBlock of an n-th-level convolution networknbAnd outputting the output characteristic diagram phi R of the b convolution module of the nth-level convolution networknb(ii) a Thereby obtaining the B-th of the 1 st convolution network1Output characteristic diagram of convolution module
The B-th of the 1 st level convolution network1Output characteristic diagram of convolution moduleAfter the processing of a convolution layer, obtaining a primary fusion image F';
step three, constructing a loss function and training a network to obtain an optimal fusion model:
step 3.1, respectively calculating image block sets { S of all modes1,…,Sm…,SMEntropy of each image block set in the image block set is obtained, and a corresponding entropy value { e }is obtained1,…,em…,eMIn which emEntropy values of a set of image blocks representing an mth modality;
step 3.2, for the entropy value { e1,…,em…,eMRespectively carrying out normalization processing to obtain image block sets (S) of all modes1,…,Sm…,SMWeight of [ omega ]1,…,ωm,…,ωMIn which ω ismWeights representing a set of image blocks of the mth modality;
and 3.3, constructing a total Loss function Loss by using the formula (1):
in the formula (1), Lssim(SmF') represents the set S of image blocks of the m-th modalitymStructural similarity with the preliminary fusion image FA loss function;
step 3.4, an optimizer is used for carrying out minimum solving on the total Loss function Loss, so that all parameters in the fusion network Transfusion are optimized, and an optimal fusion model is obtained;
step four, utilizing the optimal fusion model to perform Y-channel image { I) of all modes1,I2,…,IMProcessing and outputting a primary fusion image F'; and converting the preliminary fusion image F' into an RGB color space, thereby obtaining a final fusion image F.
2. The multi-modal medical image fusion method based on global information fusion as claimed in claim 1, wherein the n-th fusion module in the step 2.2 is processed according to the following procedures:
step 2.2.1, the characteristic diagram of the N-th convolution module output by the n-th fusion module to the M modal branch networksSplicing and leveling to obtain the size of (M H)n*Wn)×DnThe flattened feature vectors of (a); adding the flattened characteristic vector and a trainable vector with the same size to obtain a characteristic vector containing position information of the nth fusion module
Step 2.2.2, when l is 1, the ith self-attention mechanism module of the nth fusion module uses the feature vectorAfter linear mapping, three matrixes Q are obtainednl,Knl,Vnl(ii) a Recalculate Qnl,Knl,VnlMultiple head attention results in between Znl(ii) a The multi-head attention result ZnlFull connection of the ith attention mechanism module input to the nth fusion moduleIn the layer connection, the output sequence vector of the ith self-attention mechanism module of the nth fusion module is obtained
When l is 2,3, …, N, the output sequence vector of the l-1 self-attention mechanism module of the N fusion module is addedInputting the input into the first self-attention mechanism module of the nth fusion module, and obtaining the output sequence vector of the first self-attention mechanism module of the nth fusion moduleThereby obtaining the output sequence vector of the L-th self-attention mechanism module of the n-th fusion module
Priority Applications (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN202210202366.1A CN114565816B (en) | 2022-03-03 | 2022-03-03 | Multi-mode medical image fusion method based on global information fusion |
Applications Claiming Priority (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN202210202366.1A CN114565816B (en) | 2022-03-03 | 2022-03-03 | Multi-mode medical image fusion method based on global information fusion |
Publications (2)
Publication Number | Publication Date |
---|---|
CN114565816A true CN114565816A (en) | 2022-05-31 |
CN114565816B CN114565816B (en) | 2024-04-02 |
Family
ID=81717119
Family Applications (1)
Application Number | Title | Priority Date | Filing Date |
---|---|---|---|
CN202210202366.1A Active CN114565816B (en) | 2022-03-03 | 2022-03-03 | Multi-mode medical image fusion method based on global information fusion |
Country Status (1)
Country | Link |
---|---|
CN (1) | CN114565816B (en) |
Cited By (3)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN115115523A (en) * | 2022-08-26 | 2022-09-27 | 中加健康工程研究院(合肥)有限公司 | CNN and Transformer fused medical image depth information extraction method |
CN115134676A (en) * | 2022-09-01 | 2022-09-30 | 有米科技股份有限公司 | Video reconstruction method and device for audio-assisted video completion |
CN115511767A (en) * | 2022-11-07 | 2022-12-23 | 中国科学技术大学 | Self-supervised learning multi-modal image fusion method and application thereof |
Citations (4)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
KR20210091276A (en) * | 2018-11-16 | 2021-07-21 | 아리엘 에이아이, 인크. | 3D object reconstruction |
CN113469094A (en) * | 2021-07-13 | 2021-10-01 | 上海中科辰新卫星技术有限公司 | Multi-mode remote sensing data depth fusion-based earth surface coverage classification method |
US11222217B1 (en) * | 2020-08-14 | 2022-01-11 | Tsinghua University | Detection method using fusion network based on attention mechanism, and terminal device |
CN114049408A (en) * | 2021-11-15 | 2022-02-15 | 哈尔滨工业大学(深圳) | Depth network model for accelerating multi-modality MR imaging |
-
2022
- 2022-03-03 CN CN202210202366.1A patent/CN114565816B/en active Active
Patent Citations (4)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
KR20210091276A (en) * | 2018-11-16 | 2021-07-21 | 아리엘 에이아이, 인크. | 3D object reconstruction |
US11222217B1 (en) * | 2020-08-14 | 2022-01-11 | Tsinghua University | Detection method using fusion network based on attention mechanism, and terminal device |
CN113469094A (en) * | 2021-07-13 | 2021-10-01 | 上海中科辰新卫星技术有限公司 | Multi-mode remote sensing data depth fusion-based earth surface coverage classification method |
CN114049408A (en) * | 2021-11-15 | 2022-02-15 | 哈尔滨工业大学(深圳) | Depth network model for accelerating multi-modality MR imaging |
Non-Patent Citations (2)
Title |
---|
沈文祥;秦品乐;曾建潮;: "基于多级特征和混合注意力机制的室内人群检测网络", 计算机应用, no. 12 * |
洪炎佳;孟铁豹;黎浩江;刘立志;李立;徐硕?;郭圣文;: "多模态多维信息融合的鼻咽癌MR图像肿瘤深度分割方法", 浙江大学学报(工学版), no. 03 * |
Cited By (5)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN115115523A (en) * | 2022-08-26 | 2022-09-27 | 中加健康工程研究院(合肥)有限公司 | CNN and Transformer fused medical image depth information extraction method |
CN115115523B (en) * | 2022-08-26 | 2022-11-25 | 中加健康工程研究院(合肥)有限公司 | CNN and Transformer fused medical image depth information extraction method |
CN115134676A (en) * | 2022-09-01 | 2022-09-30 | 有米科技股份有限公司 | Video reconstruction method and device for audio-assisted video completion |
CN115134676B (en) * | 2022-09-01 | 2022-12-23 | 有米科技股份有限公司 | Video reconstruction method and device for audio-assisted video completion |
CN115511767A (en) * | 2022-11-07 | 2022-12-23 | 中国科学技术大学 | Self-supervised learning multi-modal image fusion method and application thereof |
Also Published As
Publication number | Publication date |
---|---|
CN114565816B (en) | 2024-04-02 |
Similar Documents
Publication | Publication Date | Title |
---|---|---|
Zhang et al. | Translating and segmenting multimodal medical volumes with cycle-and shape-consistency generative adversarial network | |
CN114565816B (en) | Multi-mode medical image fusion method based on global information fusion | |
Li et al. | VolumeNet: A lightweight parallel network for super-resolution of MR and CT volumetric data | |
CN107492071A (en) | Medical image processing method and equipment | |
Du et al. | Accelerated super-resolution MR image reconstruction via a 3D densely connected deep convolutional neural network | |
CN113506222B (en) | Multi-mode image super-resolution method based on convolutional neural network | |
WO2022121100A1 (en) | Darts network-based multi-modal medical image fusion method | |
Zhu et al. | Stacked U-shape networks with channel-wise attention for image super-resolution | |
CN110322402A (en) | Medical image super resolution ratio reconstruction method based on dense mixing attention network | |
CN111487573B (en) | Enhanced residual error cascade network model for magnetic resonance undersampling imaging | |
CN111899165A (en) | Multi-task image reconstruction convolution network model based on functional module | |
CN111598964B (en) | Quantitative magnetic susceptibility image reconstruction method based on space adaptive network | |
Singh et al. | Medical image generation using generative adversarial networks | |
CN114677263A (en) | Cross-mode conversion method and device for CT image and MRI image | |
CN115511767A (en) | Self-supervised learning multi-modal image fusion method and application thereof | |
CN112488971A (en) | Medical image fusion method for generating countermeasure network based on spatial attention mechanism and depth convolution | |
CN115457359A (en) | PET-MRI image fusion method based on adaptive countermeasure generation network | |
CN114821259A (en) | Zero-learning medical image fusion method based on twin convolutional neural network | |
Yu et al. | 3D Medical Image Segmentation based on multi-scale MPU-Net | |
Xie et al. | MRSCFusion: Joint Residual Swin Transformer and Multiscale CNN for Unsupervised Multimodal Medical Image Fusion | |
CN116309754A (en) | Brain medical image registration method and system based on local-global information collaboration | |
Li et al. | HRINet: alternative supervision network for high-resolution CT image interpolation | |
CN117036162B (en) | Residual feature attention fusion method for super-resolution of lightweight chest CT image | |
CN117649422B (en) | Training method of multi-modal image segmentation model and multi-modal image segmentation method | |
Wu et al. | Convolutional neural network with coarse-to-fine resolution fusion and residual learning structures for cross-modality image synthesis |
Legal Events
Date | Code | Title | Description |
---|---|---|---|
PB01 | Publication | ||
PB01 | Publication | ||
SE01 | Entry into force of request for substantive examination | ||
SE01 | Entry into force of request for substantive examination | ||
GR01 | Patent grant | ||
GR01 | Patent grant |